Ishan S. Kshirsagar

Latent reasoning in language models may not function as an active scratchpad during inference, but rather shapes model parameters, leading to surprising gains in robustness post-reinforcement learning.